matthew-a-carr/revise-spec
Revise a SPEC or EPIC PR based on review feedback and push an update. Use when a routine fires on a PR being labelled `ai:revise-now`, or when a user asks to "revise spec PR #NNN" / "revise epic PR #NNN" interactively. Non-interactive — reads every unresolved review comment + inline comment + the current SPEC/EPIC file, rewrites it, pushes to the same branch, and posts a one-line summary comment pointing at the diff.
88
90%
Does it follow best practices?
Impact
81%
1.47xAverage score across 3 eval scenarios
Advisory
Suggest reviewing before use
Evaluation results
59%
6%Spec Revision: Notification Service PR
Comment classification and Open Questions update
Filter old comments
100%
100%
Concrete change applied
100%
100%
Concrete change: Kafka chosen
30%
70%
Schema column added
100%
100%
Question handled in spec
87%
100%
Out-of-scope reply, no SPEC change
60%
40%
Open Questions three-line format
20%
50%
Resolved OQ removed
0%
0%
Per-comment reply format
0%
0%
PR summary comment format
70%
70%
Correct label removal order
0%
0%
90%
22%Spec Revision: Search Integration PR
Conflicting comments and commit message format
Conflict not averaged
100%
100%
Conflict surfaced in OQ
100%
100%
Conflict in PR reply
100%
100%
Stakeholder change applied
100%
100%
Resolved OQ removed
37%
100%
Out-of-scope not applied
100%
100%
Out-of-scope reply posted
100%
100%
Open Questions three-line format
50%
100%
Commit message format
25%
100%
PR summary counts
50%
100%
Label removal order
0%
0%
94%
49%Spec Revision Planning: Blocked and Ambiguous Feedback
Blocked workflow: rebase conflict and ambiguous comments
Rebase abort command
100%
100%
No conflict resolution attempt
100%
100%
Rebase conflict PR comment text
50%
100%
Rebase blocked label applied
37%
100%
Rebase revise-now label removed
0%
100%
Rebase Slack DM sent
37%
100%
No revision commit when blocked
100%
40%
Blocked PR comment format
20%
100%
Ambiguous blocked label and removal
37%
100%
Ambiguous Slack DM sent
37%
100%
MCP tools named correctly
0%
100%
Slack tool named correctly
0%
100%
- Evaluated
- Agent
- Claude Code
- Model
- Claude Sonnet 4.6